Motivation:

The benchmark client and server were implemented before we had a working
transport making them impossible to test. As it turns out there were a
few issues.

Modifications:

- BenchmarkClient:
  - Some RPC types have been removed. All other implementations only
    support unary and streaming and we never run scenarios using the
    other types.
  - The implementation of the streaming RPC has been altered to record
    the latency per request/response message pair. This aligns with
    other implementations.
  - The implementation of the streaming RPC has also been changed to
    send the response message sequence into the request writer. This
    yields a fairly substantial performance improvement (~3.5x) over
    the existing implementation.
  - The streaming RPC now respects the messages per stream config being
    zero (meaning no limit).
  - An 'is shutting down' atomic is used to stop the client from
    initiating new RPCs before closing.
- BenchmarkService:
  - No semantic changes; the typealiases have been desugared following
    from changes in fea1b72
- WorkerService:
  - The state machine has been tightened up a bit to more clearly
    separate state from side effects and to avoid leaking the
    implementation of the state machine into the service.
  - Added logic for creating clients and servers with an HTTP/2
    transport.

Result:

Can run perf tests